Effects of peripheral nonlinearity on psychometric functions for forward-masked tones a)

advertisement
Effects of peripheral nonlinearity on psychometric functions
for forward-masked tonesa)
Kim S. Schairer,b) Lance Nizami, Jason F. Reimer,c) and Walt Jesteadt
Center for Hearing Research, Boys Town National Research Hospital, 555 North 30th Street,
Omaha, Nebraska 68131
!Received 14 March 2002; revised 8 October 2002; accepted 16 December 2002"
Psychometric functions !PFs" for forward-masked tones were obtained for conditions in which
signal level was varied to estimate threshold at several masker levels !variable-signal condition", and
in which masker level was varied to estimate threshold at several signal levels !variable-masker
condition". The changes in PF slope across combinations of masker frequency, masker level, and
signal delay were explored in three experiments. In experiment 1, a 2-kHz, 10-ms tone was masked
by a 50, 70 or 90 dB SPL, 20-ms on-frequency forward masker, with signal delays of 2, 20, or 40
ms, in a variable-signal condition. PF slopes decreased in conditions where signal threshold was
high. In experiments 2 and 3, the signal was a 4-kHz, 10-ms tone, and the masker was either a 4or 2.4-kHz, 200-ms tone. In experiment 2, on-frequency maskers were presented at 30 to 90 dB SPL
in 10-dB steps and off-frequency maskers were presented at 60 to 90 dB SPL in 10-dB steps, with
signal delays of 0, 10, or 30 ms, in a variable-signal condition. PF slopes decreased as signal level
increased, and this trend was similar for on- and off-frequency maskers. In experiment 3,
variable-masker conditions with on- and off-frequency maskers and 0-ms signal delay were
presented. In general, the results were consistent with the hypothesis that peripheral nonlinearity is
reflected in the PF slopes. The data also indicate that masker level plays a role independent of signal
level, an effect that could be accounted for by assuming greater internal noise at higher stimulus
levels. © 2003 Acoustical Society of America. #DOI: 10.1121/1.1543933$
PACS numbers: 43.66.Ba, 43.66.Dc #MRL$
I. INTRODUCTION
Forward masking occurs when the threshold for detection of a brief tone or signal is elevated in the presence of a
preceding stimulus or masker. Forward masking data have
been obtained in conditions in which signal level is varied to
estimate threshold at fixed masker levels !variable-signal
condition", or conditions in which masker level is varied to
estimate threshold at fixed signal levels !variable-masker
condition". The function describing the relation between
masker level and signal level at threshold is referred to as
growth of masking !GOM". These patterns have been described in detail for variable-signal !e.g., Jesteadt et al.,
1982" and variable-masker conditions !e.g., Nelson and
Freyman, 1987", and will not be reviewed here. A recent
model of forward masking !Oxenham and Plack, 2000; Plack
and Oxenham, 1998" accounts for GOM results by assuming
a peripheral nonlinearity with a specific form that is consistent with direct basilar membrane !BM" input–output !I/O"
a"
Portions of this work were presented in ‘‘Psychometric functions for detection of tones following a forward masker,’’ 2001 MidWinter Meeting of
the Association for Research in Otolaryngology, St. Petersburg Beach, FL,
February 2001; ‘‘Psychometric functions for detection of tones following
on- and off-frequency forward maskers,’’ 141st Meeting of the Acoustical
Society of America, Chicago, IL, June 2001; and ‘‘Psychometric functions
for variable masker levels in the context of peripheral nonlinearity,’’ 2002
MidWinter Meeting of the Association for Research in Otolaryngology, St.
Petersburg Beach, FL, January 2002.
b"
Electronic mail: schairerk@boystown.org
c"
Current address: Department of Psychology, California State University,
San Bernardino, 5500 University Parkway, San Bernardino, CA 924072397.
1560
J. Acoust. Soc. Am. 113 (3), March 2003
recordings !Rhode, 1971; Robles et al., 1986; Ruggero et al.,
1997" and with estimates from auditory nerve rate-intensity
functions !Yates et al., 1990". The BM I/O function is described as having a compressive, nonlinear growth when the
signal frequency is equal to the characteristic frequency
!CF", which is analogous to the case in which the masker and
signal are the same frequency !on-frequency masker", and
more linear growth when the signal frequency is different
from CF, which is analogous to the case in which the masker
is a different frequency than the signal !off-frequency
masker". The nonlinearity at CF can be characterized by two
straight lines in log–log coordinates, with a low-level linear
region !i.e., slope of 1.0", and a moderate-level compressive
region with a slope of around 0.2 above some breakpoint
!Yates et al., 1990".
The Plack and Oxenham !1998" model assumes that a
sliding temporal integration window receives the output of
the peripheral nonlinearity, and serves to overlap the masker
and signal. In a two-interval forced choice task !2IFC", the
output of the temporal window is compared for the two intervals, and the subject is assumed to vote for the interval
with the greater output. Multiplicative internal noise limits
performance and is assumed to be uniform across all conditions. The current paper is concerned only with the peripheral nonlinearity component of the model. Therefore, the remaining components of the model and their underlying
processes will not be discussed further here.
The Plack and Oxenham !1998" model describes the relationship between signal level and masker level when both
of these quantities are changing, as in GOM functions. The
0001-4966/2003/113(3)/1560/14/$19.00
© 2003 Acoustical Society of America
FIG. 1. Example of a compressive nonlinearity similar to the nonlinearity
assumed in the model of forward masking. For conditions that produce low
signal level at masked threshold, the signal passes through the more linear
portion of the function as its level varies during the threshold estimation
procedure. Changes in signal level at the input of the function will result in
similar changes in signal level at the output of the function, and a smaller
range of input levels will be presented in establishing threshold. The underlying distribution of input signal levels in this case is represented by the
probability density function !PDF" on the left along the X axis, and the
corresponding distribution for signal levels at the output is represented by
the lower PDF on the Y axis. PF slopes, which reflect the standard deviation
of these underlying PDFs, would be steep for low-threshold conditions because the standard deviation of the distribution on the X axis would be
small. In conditions that produce higher signal levels at masked threshold,
the signal will pass through the more compressive region of the function,
and will have to change by a greater amount !PDF on the right along the X
axis" to produce the same change in signal level at the output of the nonlinearity !upper PDF along the Y axis". In this case, the PF slope would be
more shallow !i.e., the underlying PDF would have a larger standard deviation". Internal noise is assumed to be uniform across conditions, and is
reflected in this figure by similar standard deviations for the upper and lower
PDFs on the Y axis.
dynamic nature of this relationship has tended to obscure the
role of the peripheral nonlinearity and the model has provided valuable insights concerning the effects of specific
changes in stimulus parameters or of changes in the nonlinearity itself !e.g., Oxenham and Plack, 1997". Psychometric
functions !PFs", which describe the relationship between
stimulus level and percent correct in a given condition, may
provide a means of exploring the effect of the nonlinearity in
a simpler framework in which only one quantity, either signal or masker level, is varied. In particular, the slope of the
PF might be expected to provide a direct measure of the
degree of nonlinearity at the signal place on the BM in the
range of intensities traversed by the signal or masker.
For example, in variable-signal conditions, the combination of a moderate masker level and a longer signal delay
produces a low signal level at masked threshold. In this case,
the signal passes through the more linear portion of the BM
I/O function as its level varies. Thus, changes in signal level
at the input of the cochlear nonlinearity will result in similar
changes in signal level at the output of the nonlinearity, and
a smaller range of input signal levels will be presented in
establishing threshold !Fig. 1". The underlying distribution of
input signal levels in this case is represented by the probability density function !PDF" on the left along the X axis, and
the corresponding distribution for signal levels at the output
of the nonlinearity is represented by the lower PDF on the Y
axis. PF slope is related to the PDF width such that the wider
the PDF !i.e., the greater the standard deviation of the underJ. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
lying distribution", the shallower the PF slope. Thus, associated PFs for low masked-threshold conditions would be
steep because the standard deviation of the distribution on
the X axis would be small. In conditions that produce higher
signal levels at masked threshold, such as with shorter signal
delay and/or higher masker level, the signal will pass
through the more compressive region of the BM I/O function. The signal level at the input of the nonlinearity will
have to change by a greater amount to produce the same
change in signal level at the output of the nonlinearity, thus
the PF slope would be more shallow. This relationship is
illustrated by the wider PDF on the right along the X axis,
and the corresponding upper PDF on the Y axis. Note that the
range of the PF at the output of the nonlinearity is specified
by internal noise, which is assumed to be uniform across
conditions !Plack and Oxenham, 1998". This is illustrated by
similar standard deviations for the upper and lower PDFs on
the Y axis. We also would predict that PF slopes should be
similar for on- and off-frequency maskers in the variablesignal case if the signal levels at threshold are similar for the
two conditions. When signal level is varied in order to determine threshold at a fixed masker level, the effect of the nonlinearity should be independent of the properties of the
masker.
In the current experiments 1 and 2, variable-signal conditions were presented that produced a range of signal levels
at masked threshold for on- and off-frequency maskers.
Mean thresholds from the on-frequency masker case in experiment 2 were used to estimate the parameters of the nonlinearity as described by Plack and Oxenham !1998" using
data analysis software provided by Plack. The purpose was
to demonstrate the ability of the Plack and Oxenham model
to predict our own masked thresholds before extrapolating
the model predictions to PFs. Next, PF slopes were estimated
and compared across conditions. It was hypothesized that PF
slope would decrease as a function of signal level, and that
this trend would be similar for on- and off-frequency masker
cases.
Although variable-masker conditions have been used
frequently in forward masking studies !Nelson et al., 2001,
1990; Nelson and Freyman, 1987, 1984; Oxenham and
Plack, 1997", PFs have not been reported for those conditions. When the masker level is varied to determine threshold
at a fixed signal level, the masker travels through the nonlinearity at the place of the signal, but the growth of the masker
at the place of the signal is different for on- and offfrequency maskers. This is because the growth of the onfrequency masker will be influenced by the compressive
nonlinearity at the place of the signal, whereas the growth of
the off-frequency masker will be linear at the place of the
signal. Thus, we predicted that PF slopes should be different
for on- and off-frequency maskers in variable-masker conditions. Masker levels were varied to find threshold at several
fixed signal levels for on- and off-frequency maskers in experiment 3 and PFs were fitted to those data.
It should be emphasized that the purpose of the current
paper is to demonstrate that the peripheral nonlinearity has
an effect on PF slopes, and not to identify a specific form of
the underlying nonlinear function. More recent models that
Schairer et al.: Peripheral nonlinearity and psychometric functions
1561
incorporate a compressive nonlinearity have assumed a continuously level-dependent function, such as that employed in
Glasberg and Moore !2000", rather than the two-straight-line
approximation assumed in the Plack and Oxenham !1998"
model. We focus on the latter model because its properties
allow for specific predictions about forward-masked thresholds and associated PFs, and because we have used select
conditions from Plack and Oxenham !1998" and Oxenham
and Plack !2000" in our experiments 1 and 2.
All of the PFs reported here have been reconstructed
from trial-by-trial data obtained with an adaptive procedure
using data collection and analysis procedures described and
tested by Dai !1995". These procedures provided an efficient
method of obtaining PFs, while allowing us to obtain threshold estimates using a procedure comparable to that used by
Plack and Oxenham !1998". As a result of using an adaptive
procedure and then fitting a PF to the data, there are two
‘‘threshold’’ estimations described in the Results section for
each experiment. The thresholds estimated from the adaptive
procedures are presented first !i.e., the GOM functions", then
the PF data are presented, which include slope and threshold
from the PF fitting procedures. Where PF slopes are compared to threshold, ‘‘threshold’’ in this case refers to the values obtained from the PF fitting procedure, although the
good agreement between these measures makes them interchangeable.
II. EXPERIMENT 1
Variable-signal conditions were presented for combinations of three on-frequency masker levels and three signal
delays. PFs were reconstructed for each subject in each condition. If PF slope is determined by the degree of peripheral
nonlinearity, and only by the degree of peripheral nonlinearity, then slopes should decrease as masked threshold for the
signal increases.
A. Subjects
Three subjects, ages 20 to 41, participated in experiment
1 !S1–S3". Two of the subjects were the first and second
authors. The other subject was a college student from a local
university who was paid for his participation. This subject
was informed of the procedures and purpose of the study,
and signed a consent form prior to participation. Hearing for
the authors was screened at 0.5, 1.0, 2.0, and 4.0 kHz in both
ears, using laboratory stimulus generation and threshold
measures as described in the following sections. Thresholds
were at or better than 20 dB SPL at each test frequency,
bilaterally. Hearing for the third subject was screened using
standard clinical procedures as part of a research project for
another laboratory at this facility. Results from that test indicated thresholds were at or better than 15 dB HL at octave
and interoctave test frequencies from 0.25 to 8.0 kHz, bilaterally.
B. Stimuli
The variable-signal conditions in the current experiment
1 were a subset of the conditions used by Plack and Oxenham !1998" in their experiment 1. The signal was a 2-kHz,
1562
J. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
10-ms tone !5-ms rise/fall" shaped with a cosine-squared envelope. The forward masker was a 2-kHz, 20-ms tone !2-ms
rise/fall" shaped with a cosine-squared envelope, presented at
50, 70, or 90 dB SPL. Each masker level was presented in
conditions in which the signal delay !measured at 0-voltage
points" was 2, 20, or 40 ms. Threshold for the signal in quiet
also was determined for each subject.
The masker and signal were generated digitally at 50
kHz !TDT AP2", and routed through separate 16-bit D/A
converters !TDT DD1", then to an anti-aliasing filter !FT6"
with a 20-kHz low-pass cutoff. The filtered stimuli were
routed to programmable attenuators !PA4", mixed !SM3",
then presented through headphone buffers !HB6" to the left
ear of each participant through circumaural headphones
!Sennheiser HD 250 II". Additional attenuation of up to 30
dB was available from resistor pads located in each booth.
The headphones were plugged directly into boxes that contained these resistor pads.
C. Threshold determination
Thresholds for the signal in quiet and in the various
experimental conditions were obtained using a 2IFC adaptive
procedure, in which the level of the signal was decreased
after two correct responses and increased after one incorrect
response !Levitt, 1971". The step size was 8 dB until the
fourth reversal, after which the step size was 4 dB. The step
sizes were selected to reduce the number of levels visited by
the adaptive procedure and increase the number of trials per
level, an approach that Dai !1995" found to result in more
stable estimates of PF parameters. Subjects responded using
hand-held microterminals !Intelligent Instrumentation
CTM280B". A trial consisted of a 500-ms warning interval
!indicated by a ‘‘* ’’ on the microterminal", two observation
intervals !indicated by a ‘‘1’’ and ‘‘2’’" separated by 500 ms,
an answer interval !indicated by a ‘‘?’’", and a 300-ms feedback interval in which the correct answer was displayed.
There was a 500-ms interval between the trials, and 50 trials
were presented in each block. Threshold for a block was
calculated as the mean of the signal levels of all reversals
after the fourth reversal. Conditions were presented in counterbalanced order. Subjects completed a minimum of 13 50trial blocks for each of the nine conditions !three signal levels ! three masker levels". Data were collected in 2-h
sessions, and subjects were given breaks periodically during
the sessions. Mean threshold for each subject in each condition was calculated using thresholds from the last 12 blocks
of trials.
D. Psychometric functions
PFs have been analyzed in many different coordinates,
particularly in the literature on intensity discrimination !e.g.,
Buus and Florentine, 1991; Moore et al., 1999". In the literature on detection, the most common approach is to assume
that the PF can be described by
d ! "m ! I " k ,
where the special cases of amplitude detection and energy
detection are represented by k"0.5 and k"1, respectively
Schairer et al.: Peripheral nonlinearity and psychometric functions
FIG. 2. Signal levels at masked threshold as a function of masker level, with
signal delay as the parameter, for the mean across subjects for experiment 1.
The forward masker was a 2-kHz, 20-ms tone presented at 50, 70, or 90 dB
SPL and the signal was a 2-kHz, 10-ms tone presented at 2- !triangles", 20!squares", or 40-ms !circles" signal delay in a variable-signal condition.
Error bars represent plus and minus one standard deviation. The dashed line
represents mean threshold for the signal in quiet !QT". Thresholds are low
and slopes of GOM for the 20- and 40-ms delay conditions are shallow.
Thresholds are higher in the 2-ms delay condition, and the slope of GOM is
steeper for all subjects.
!Egan et al., 1969". Other authors have assumed an underlying function of this form when simulating recovery of PF
parameters from adaptive-tracking data !Dai, 1995; Leek
et al., 1992". Dai !1995" used a modified form of the equation
d ! " ! I/I o " k
where I o "(I/m) k in the Egan et al. equation. In Dai’s version, 10 log(Io) represents the level required for d ! "1,
whereas values of m in the original equation have no meaning independent of k. In Dai’s approach, I o and k are varied
to minimize the weighted squared deviation between obtained and estimated values of percent correct, a % 2 measure.
Weights reflect the standard errors associated with the estimated percent correct values. The exponent k can be treated
as a measure of the slope of the PF, given that 10 log d!
"k*10 log(I/Io). We have used Dai’s version of the formula
in the data analyses reported here.
The trial-by-trial data were pooled across the last 12
blocks to obtain estimates of percent correct as a function of
signal level for each subject in each condition. Signal levels
that were presented on 30 trials or less and/or produced less
FIG. 3. Mean PFs for all nine conditions in experiment 1. The 50-dB
masker conditions are represented by the solid lines, 70-dB maskers by
dotted lines, and 90-dB maskers by dashed lines. Signal delay is not directly
identified in this figure. As the signal levels increase, the slopes of the PFs
decrease, indicating that presentation of a wider range of signal levels was
required to determine threshold at higher levels.
than 50% correct were not included in the fits.1 The Dai
fitting procedure provided the threshold and the slope of the
PF for each condition for each subject. These parameters
were averaged across the three subjects to obtain grand average PFs in each condition. The average across subjects
included only those PF fits that had an r 2 of 0.50 or greater.
E. Results
Mean signal levels at threshold across subjects as a function of masker level, with signal delay as the parameter, are
displayed in Fig. 2. Error bars represent the standard deviation of thresholds across subjects. Masked thresholds for the
20- and 40-ms delay conditions are low, slopes of GOM !i.e.,
the rate at which masked threshold increases as a function of
masker level" are shallow, and the two functions are similar.
Masked thresholds are much higher in the 2-ms delay condition, and the slope of GOM is much steeper.
Although Dai’s maximum-likelihood procedure for fitting PFs minimizes % 2 , unweighted r 2 values provide a more
interpretable measure of goodness of fit. Table I provides a
summary of the PF thresholds and slopes, and r 2 values for
the individual PF fits. All but one fit met the criteria of r 2
#0.50 for inclusion in the mean data !S1, 50-dB masker,
2-ms signal delay". Mean PFs for all nine conditions are
shown in Fig. 3. Overall, as the signal levels increase, the
slopes of the PFs decrease. Thus, presentation of a wider
TABLE I. PF thresholds and slopes, and r 2 values for PF fits for each subject in each condition in experiment 1.
Subject
S1
S2
S3
2-ms signal delay
20-ms signal delay
Masker
level
Threshold
Slope
r
50
70
90
50
70
90
50
70
90
17.98
30.46
45.96
39.30
48.06
56.72
39.45
48.65
65.84
0.10
0.36
0.28
0.67
0.60
0.36
0.39
0.20
0.28
0.40
0.94
0.76
0.96
0.97
0.92
0.93
0.83
0.82
J. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
2
40-ms signal delay
2
Threshold
Slope
r
12.24
13.67
19.16
29.44
32.27
32.59
24.38
27.54
32.54
0.65
0.37
0.39
0.58
0.54
0.69
0.58
0.74
0.54
0.97
0.92
0.89
0.95
0.88
0.99
0.93
0.80
0.94
Threshold
Slope
r2
9.63
8.99
9.16
30.79
31.33
28.71
22.45
23.21
26.50
0.84
0.60
0.65
0.56
1.11
1.03
0.92
0.90
0.94
0.95
0.96
0.88
0.90
0.99
0.99
0.99
0.89
0.96
Schairer et al.: Peripheral nonlinearity and psychometric functions
1563
III. EXPERIMENT 2
Variable-signal conditions were presented using three
shorter signal delays and a longer duration masker to produce more masking than in experiment 1. Off-frequency forward maskers also were included. The stimuli and signal
delays were the same as those in experiment 1 of Oxenham
and Plack !2000". Software supplied by Plack was used to
estimate the nonlinearity parameters from the mean data
from on-frequency masker conditions. It was predicted that
PF slopes would decrease as a function of signal level at
masked threshold, and that this trend would be similar for
on- and off-frequency masker conditions.
FIG. 4. Mean slopes of PFs as a function of mean threshold in dB SPL
across subjects for all masker levels and signal delays in experiment 1. The
filled circle represents the PF threshold and slope for the quiet threshold
condition !QT". The r 2 value of 0.72 is from the linear regression fit. Note
that this r 2 is associated with the regression line for slope versus threshold,
not individual PF fits.
range of signal levels is required to determine threshold at
higher levels. In Fig. 4, the mean slope of PFs across subjects is shown as a function of mean threshold for the nine
masker level and signal delay combinations. PF slope decreases as masked threshold increases, and the relation is
well described by a straight line. However, the data points
for the two longer signal delays are not well separated, either
in terms of slope or threshold. Shorter signal delays may
have produced a better spread of masked thresholds and a
more appropriate set of data for regression analysis.
Individual PF fits to the data for three of the nine conditions are shown in Fig. 5. As stated previously, the two
longer signal delays produced mean masked thresholds that
were very close together !see Fig. 2", and this is reflected in
the PFs in the lower signal level range for individual subjects. The data points, particularly in the middle of the percent correct range, are well represented by the fits. However,
values closer to 50% correct are weighted less in the fitting
procedure, and sometimes are not well represented by the
fitted line.
A. Subjects
Three males and four females from 19 to 35 years of age
participated in experiment 2 !S4 –S10". Two of the males and
one of the females were tested simultaneously in one group,
and the remaining subjects were tested together in a second
group. Each participant was informed of the procedures and
purpose of the study, and signed a consent form prior to
participation. All participants were college students from a
local university, and none had previous experience with the
task. They were paid for their participation. Hearing was
screened at 0.5, 2.0, and 4.0 kHz in both ears using the same
procedures as described for the authors in experiment 1.
Thresholds for all participants were at or better than 20 dB
SPL at each test frequency, bilaterally.
B. Stimuli
The signal was a 4-kHz, 10-ms tone !5-ms rise/fall"
shaped with a cosine-squared envelope, and presented at delays of 0, 10, or 30 ms. The on-frequency forward masker
was a 200-ms tone, presented at 30 to 90 dB SPL in 10-dB
steps in different blocks. The off-frequency forward masker
was a 2.4-kHz, 200-ms tone presented at 60 to 90 dB SPL in
10-dB steps in different blocks. Both maskers were shaped
with 2-ms cosine-squared rise/fall envelopes. Stimulus generation and delivery was the same as for experiment 1, with
FIG. 5. Individual PF fits for select variable-signal conditions for experiment 1. The open symbols represent the actual data points, and the solid lines represent
the fits to the points. The parameter and r 2 values are listed in Table I. Figure 2 shows that the two longer signal delays produced mean masked thresholds
that were very close together, and this is reflected in the PFs for individual subjects. The fits to the data points are good, particularly in the middle of the
percent correct range. Values at the lower end of the percent correct scale tend to be weighted less in the fitting procedure, and sometimes are not well
represented by the fitted line. In these individual fits, particularly for S3, it can be seen that PF slopes tend to decrease as signal level increases.
1564
J. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
Schairer et al.: Peripheral nonlinearity and psychometric functions
the exception that multiple sets of attenuators, mixers, and
headphone buffers were used to enable simultaneous testing
of up to four subjects.
C. Threshold determination
Thresholds were determined using the same procedures
as experiment 1, with the exception that step sizes were all 4
dB. The first group !S5, S7, and S9" completed 12 50-trial
blocks of each condition; the second group !S4, S6, S8, S10"
completed eight blocks per condition. Data from all subjects
were combined and treated as one group for subsequent
analyses.
D. Data analysis
Plack and Oxenham !1998" used temporal window parameters established by Oxenham and Moore !1994" and
nonlinearity parameters established by Oxenham and Plack
!1997" to fit their data. After adjusting the breakpoint of the
nonlinearity from 45 dB !used by Oxenham and Plack, 1997"
to 35 dB, they varied only the efficiency parameter, k, the
threshold-level difference at the output of the temporal window required for detection of the signal !Plack and Oxenham, 1998". The model assumes that k is constant across
conditions. In a later study, Oxenham !2001" varied the
lower !more linear" and upper !more compressive" slope of
the nonlinearity as well as window parameters to obtain an
optimum fit, but allowed only a few parameters to vary at
once. In fitting the current data, we have taken a similar
approach. Using the standard window parameters, two successive iterations were performed to arrive at best estimates
of the nonlinearity parameters based on the group mean
adaptive thresholds in the on-frequency conditions.2 First,
lower slope was fixed at 0.78 and k, breakpoint, and upper
slope were allowed to vary until a least squares fit was obtained. The resultant values for k, breakpoint, and upper
slope were then fixed, and lower slope was allowed to vary
until a least squares fit was obtained.
PFs were calculated for each subject in each condition
using the Dai !1995" procedure. Signal levels with 30 trials
or less and/or less than 50% correct were excluded from the
fits. Of the 231 PFs !excluding those for quiet threshold" for
seven listeners and 33 conditions, 218 were used in further
analyses. Twelve PFs were excluded because of poor fits
(r 2 $0.50), and one was excluded because only one signal
level remained after the trim procedure. Poor fits typically
resulted from nonmonotonic functions. Correlations were
calculated using SPSS to examine the relationship between
mean PF slope and mean threshold from the PF fits, masker
level, signal delay, and masker frequency.
E. Results
Mean signal levels at masked threshold !across replications, then across subjects" are shown as a function of
masker level with signal delay as the parameter in Fig. 6. In
the on-frequency masker conditions, the shorter signal delays
and longer masker duration used in experiment 2 produced
higher signal levels at masked threshold in comparison to
experiment 1. In both the on- and off-frequency conditions,
J. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
FIG. 6. Mean signal levels at threshold as a function of masker level with
signal delay as the parameter for on-frequency !left panel" and off-frequency
!right panel" masker conditions for experiment 2. The on-frequency forward
masker was a 4-kHz, 200-ms tone presented at 30 to 90 dB SPL in 10-dB
steps, and the off-frequency forward masker was a 2.4-kHz tone presented at
60 to 90 dB SPL in 10-dB steps. The signal was a 4-kHz, 10-ms tone
presented at 0- !triangles", 10- !squares", or 30-ms !circles" signal delay in a
variable-signal condition. Error bars represent plus and minus one standard
deviation across subjects. The solid lines in the left panel for the onfrequency masker conditions represent the model fit to the mean data obtained with the specified parameters. The dotted line in the right panel represents threshold for the signal in quiet.
signal level at masked threshold increases with increases in
masker level. However, the slopes of GOM are much steeper
for the off-frequency maskers, and remain steep as the delay
increases, whereas the slopes of GOM tend to decrease with
an increase in signal delay for the on-frequency masker case.
The final fits of the Plack and Oxenham model to the
mean on-frequency masking data also are shown in the left
panel of Fig. 6. The rms error was 1.30 dB. The lower slope
that produced the best fit was 0.62 dB/dB. This value is
lower than the 0.78 dB/dB slope that Plack and Oxenham
use, and physiological data suggest a value closer to 1.0. The
upper slope of 0.12 dB/dB and breakpoint of 39 dB agree
more closely with the values of 0.16 dB/dB and 35 dB assumed by Plack and Oxenham. The efficiency factor, k, was
2.58 dB.
The PF thresholds and slopes estimated from the Dai
procedure and the corresponding r 2 values for a subset of the
conditions are summarized in Table II. Because there is a
large amount of data, only the on-frequency, 30-, 60-, and
90-dB masker conditions, and the off-frequency 60- and
90-dB masker conditions for each signal delay are presented
to demonstrate the range of parameters and goodness of fit to
individual subject data. PF fits to the individual data for a
sample of low and high masked threshold conditions are presented in Fig. 7. PF fits whose parameters are included in
Fig. 8 and in other summaries are shown by solid lines, and
PF fits that are excluded are shown by dotted lines. These
excluded conditions are the on-frequency, 60-dB masker,
30-ms signal delay condition for subject S7, and the onfrequency, 90-dB masker, 0-ms signal delay condition for
subjects S6, S8, and S10.
Figure 8 shows the PF slopes plotted as a function of the
PF thresholds for individual subjects, and the mean PF slope
as a function of mean PF threshold across subjects. Note that
subject S8 has extreme slope values !i.e., greater than 3.00"
in six conditions, and they are not plotted here in order to
keep the Y-axis scale in a range that is appropriate for the
Schairer et al.: Peripheral nonlinearity and psychometric functions
1565
TABLE II. PF thresholds and slopes, and r 2 values for PF fits for each subject for select conditions in experiment 2.
0-ms signal delay
10-ms signal delay
Subject
Masker
frequency
Masker
level
Threshold
Slope
r
S4
On
30
60
90
60
90
30
60
90
60
90
30
60
90
60
90
30
60
90
60
90
30
60
90
60
90
30
60
90
60
90
30
60
90
60
90
34.60
54.33
82.76
25.12
81.79
34.68
56.14
85.80
27.50
69.06
34.68
54.22
76.30
29.65
64.84
21.41
57.19
88.17
34.16
77.57
31.20
58.72
86.33
29.57
78.63
27.80
61.39
83.73
28.82
76.65
37.53
58.61
85.52
34.66
81.94
0.97
0.88
0.24
0.91
0.48
0.91
1.19
0.47
1.06
0.38
0.98
0.60
0.10
1.47
0.53
0.16
0.35
0.30
0.29
0.26
2.23
2.02
0.24
0.66
0.94
0.26
0.66
0.25
1.07
0.38
0.62
0.45
0.06
0.35
0.63
0.86
0.97
0.64
0.99
0.82
0.75
0.97
0.83
0.98
0.93
0.93
0.89
0.37
0.99
0.95
0.21
0.53
0.96
0.52
0.62
0.98
0.97
0.41
0.86
0.96
0.99
0.85
0.68
0.95
0.88
0.98
0.99
0.18
0.79
0.92
Off
S5
On
Off
S6
On
Off
S7
On
Off
S8
On
Off
S9
On
Off
S10
On
Off
2
30-ms signal delay
2
Threshold
Slope
r
29.84
41.08
63.28
23.19
63.10
32.92
47.62
58.02
29.08
57.45
31.34
42.00
50.42
26.96
52.31
28.05
36.83
49.23
29.91
54.54
29.56
44.66
66.60
29.57
71.34
33.77
50.07
69.65
28.04
67.67
33.73
52.53
87.78
25.33
80.14
0.82
0.65
0.24
1.10
0.35
1.30
0.92
0.22
1.45
0.33
0.98
0.33
0.64
2.12
1.04
2.23
0.79
0.13
0.35
0.43
4.19
0.75
0.25
1.29
0.27
0.88
0.57
0.19
0.66
0.30
0.54
0.40
0.24
0.48
0.33
0.90
0.99
0.80
0.85
0.92
1.00
0.99
0.80
1.00
0.81
0.92
0.97
0.97
0.99
0.99
0.99
0.91
0.61
0.85
0.83
1.00
0.86
0.95
1.00
0.66
0.89
0.96
0.84
0.88
0.89
0.93
0.88
0.66
0.98
0.83
Threshold
Slope
r2
24.76
30.22
49.76
20.85
43.37
30.67
36.41
44.82
28.99
44.39
30.20
34.98
41.45
24.57
One-point
29.32
30.05
36.55
26.15
45.03
27.15
34.06
44.52
27.59
54.12
31.27
39.27
55.83
27.23
51.82
29.06
42.27
60.08
25.35
81.11
1.41
0.75
0.38
1.33
0.29
0.49
1.13
0.96
1.17
0.61
0.76
1.03
0.64
0.99
function
0.28
0.08
0.70
0.72
0.30
4.20
1.72
0.23
1.19
0.45
0.61
0.51
0.27
1.02
0.26
0.33
0.28
0.42
0.53
0.56
0.99
0.93
0.46
0.98
0.98
0.54
1.00
0.96
0.98
0.83
0.91
0.95
0.87
0.92
0.58
%0.13
0.89
0.94
0.70
0.98
1.00
0.74
0.95
0.68
0.87
0.92
0.88
0.69
0.69
0.94
0.91
0.83
0.91
0.98
FIG. 7. PF fits to the individual data for a sample of low threshold !60-dB masker, 30-ms signal delay, represented by triangles" and high threshold conditions
!90-dB masker, 0-ms signal delay, represented by squares" for on-frequency !open symbols" and off-frequency masker cases !filled symbols" for experiment
2. PF fits that were excluded from summaries due to r 2 $0.50 are shown by dotted lines.
1566
J. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
Schairer et al.: Peripheral nonlinearity and psychometric functions
FIG. 8. Slopes of PFs plotted as a function of threshold for individual subjects, and the mean PF slope as a function of mean threshold for experiment 2. Onand off-frequency masker conditions are represented by open and closed symbols, respectively. The 0-ms signal delay conditions are represented by triangles,
10-ms delay by squares, and 30-ms delay by circles. Masker level is not directly identified in this plot. Threshold for the signal in quiet is represented by the
star. Slopes of PFs decrease as signal threshold increases, and there does not appear to be a systematic difference between masker frequency conditions.
However, there does appear to be a large amount of variability across subjects. Subject S10 has relatively uniform PF slopes across conditions. Also, subjects
S4, S5, and S8 have higher signal thresholds in the on-frequency, 0-ms condition !open triangles" without a corresponding decrease in the slopes of the PFs.
It should be noted that subject S8 has extreme slope values !i.e., #3.00" for five of the low-level, on-frequency masker conditions, and for the threshold in
quiet condition. These data points are not plotted because the Y axis would have to be expanded to accommodate them, making the trends in the data for other
subjects difficult to distinguish.
remaining subjects. For this subject, the slope values in the
40-dB, on-frequency masker conditions are 4.13, 3.06, and
3.55 in the 0-, 10-, and 30-ms signal delay conditions, respectively. Slope values in the 30-dB, on-frequency masker
condition are 4.19 and 4.20 for the 10- and 30-ms signal
delay conditions, respectively, and the slope value for threshold for the signal in quiet is 3.02. All of these conditions
have associated PF thresholds that are below 41 dB SPL. In
general, PF slopes decrease as signal level at threshold increases as observed in experiment 1. Further, there does not
appear to be a systematic difference between on- and offfrequency masker conditions. However, there does appear to
be a large amount of variability across subjects. Subject S10,
for example, has relatively uniform PF slopes across conditions. Also, although the mean data do not indicate a systematic difference among signal delays and masker frequencies,
subjects S4, S5, and S8 have higher masked thresholds in the
on-frequency, 0-ms condition !open triangles" without a corresponding decrease in the slopes of the PFs.
The correlation of mean PF slope with mean PF threshold in Fig. 8 is %0.69. We also evaluated the correlation of
mean PF slope with the stimulus parameters. The correlation
is %0.89 with masker level, 0.06 with signal delay, and 0.18
with masker frequency. Only the correlations between PF
slope and PF threshold and between PF slope and masker
level are significant (p$0.05). PF threshold also is significantly correlated with masker level !0.66" and with signal
delay !%0.49". The higher correlation of PF slope with
masker level than with signal level is unexpected, and may
be due in part to results obtained for the 0-ms signal delay,
on-frequency masker condition. When subjects S4, S5, and
S8 are excluded from the mean data for the 0-ms signal
J. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
delay, on-frequency masker condition, the correlation of
mean PF slope is %0.77 with mean PF threshold, %0.78 with
masker level, 0.28 with signal delay, and 0.49 with masker
frequency.
Given the form of the nonlinearity assumed by Plack
and Oxenham !1998", described by two straight lines in dB
coordinates, all PFs with signal levels falling only on the
lower slope should be uniformly steep, while all PFs with
signal levels falling only on the upper slope should be uniformly shallow. The pattern of results in Fig. 8 suggests a
smooth nonlinearity that is continuously decreasing in slope
and is consistent with recent models that incorporate smooth
nonlinear functions !e.g., Glasberg and Moore, 2000" and
with data that provide independent support for those models
!e.g., Nelson et al., 2001".
IV. EXPERIMENT 3
Variable-masker conditions were presented for on- and
off-frequency maskers at one signal delay. In this case, it was
predicted that the slopes of PFs for the on-frequency maskers
would be shallow at higher masker levels, and that slopes of
PFs for the off-frequency maskers would be steep across
masker levels.
A. Subjects and procedures
Three females and one male from experiment 2 participated in experiment 3 !S4, S6, S8, and S10". The signal was
a 4-kHz, 10-ms tone !5-ms rise/fall" shaped with a cosinesquared envelope, presented 0 ms after the offset of the
masker. The on-frequency forward masker was a 4-kHz,
200-ms tone. The off-frequency forward masker was at 2.4
Schairer et al.: Peripheral nonlinearity and psychometric functions
1567
kHz. Maskers were shaped with 2-ms rise/fall cosine-squared
envelopes. Stimulus generation and delivery were the same
as for experiments 1 and 2.
Initially, signal levels were fixed at 30 and 50 dB SPL
for both on- and off-frequency maskers. Thresholds were obtained using a 2IFC adaptive procedure in which the level of
the masker was decreased after two correct responses and
increased after one incorrect response. Step sizes were 4 dB.
Threshold for a block was calculated as the mean of the
masker levels of all reversals after the fourth reversal. Sixteen blocks of each condition were presented in a counterbalanced order. There was an extended break after completing data collection for these four conditions. In the process of
data analysis, an error in the stimulus file was discovered for
the off-frequency masker, 30-dB SPL signal level condition.
All blocks in this condition were repeated for all subjects,
along with four more blocks of the other three conditions as
a check for consistent thresholds. The first four blocks that
were collected previously in these conditions were excluded
from further analysis.
The initial data for masker levels at threshold for fixed
signal-level conditions of 30 and 50 dB were used to select
signal levels that would produce an overlap in the range of
masker levels at threshold for on- and off-frequency conditions. The final data set included threshold for the signal in
quiet, and masked thresholds for fixed signal levels of 30, 50,
57, and 63 dB SPL for the on-frequency masker case, and 25,
30, 40, and 50 dB SPL for the off-frequency masker case.
Twelve 50-trial blocks were completed for threshold in quiet,
and 16 50-trial blocks were completed in the remaining conditions, with the exception of the off-frequency masker in the
25-dB SPL signal level condition. S6’s threshold for the signal in quiet was 24.9 dB SPL, and she reported difficulty
hearing the signal and masker in the 25-dB SPL fixed-signal
condition. The condition was terminated after eight blocks
for all subjects.
Subjects occasionally produced masked thresholds that
clearly deviated !e.g., by more than 8 dB" from the rest of
their masked thresholds in a particular condition. These data
were excluded, and runs for the corresponding conditions
were repeated. In one instance, a subject produced unusually
variable thresholds for the 63-dB SPL fixed-signal, onfrequency masker condition !8.95 dB standard deviation
across the 16 thresholds", which were all collected during
one test session. These initial thresholds were discarded, and
16 more repetitions were collected. This second set of thresholds had a lower standard deviation, and were used for further data analysis.
PFs were calculated using the Dai !1995" method. However, because PFs in this case described percent correct as a
function of masker level, percent correct increased as masker
level decreased. PFs were fitted for individual subjects, and
only those fits that produced r 2 values of #0.50 were included in further analyses. With this criterion, PFs were excluded for the off-frequency masker, 25 dB SPL signal condition for subjects S6 and S10. Correlations were calculated
for mean PF slope as a function of mean PF threshold for onand off-frequency masker conditions.
1568
J. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
FIG. 9. Mean masker levels at threshold across subjects as a function of
signal level for experiment 3. The on-frequency forward masker !open symbols" was a 4-kHz, 200-ms tone and the off-frequency forward masker
!filled symbols" was a 2.4-kHz tone presented in a variable-masker condition. The signal was a 4-kHz, 10-ms tone presented at 0-ms signal delay.
Mean threshold for the signal in quiet is represented by the dotted line. The
signal was fixed in different conditions at 30, 50, 57, and 63 dB SPL for the
on-frequency masker case, and at 25, 30, 40, and 50 dB SPL for the offfrequency masker case. Masker level at threshold is much higher in the
off-frequency masker case than for the on-frequency case, particularly for
the lower signal levels. The slope of GOM is more shallow for the offfrequency case. These trends agree with the literature.
B. Results
Mean growth of masking !GOM" functions in the
variable-masker conditions are shown in Fig. 9. The GOM in
the variable-masker conditions was steeper for on-frequency
maskers than for off-frequency maskers. The PF thresholds,
slopes, and r 2 values for each subject in each condition are
summarized in Table III. The correlation of mean PF slope
and mean PF masker level at masked threshold was %0.88
for the on-frequency masker case and %0.87 for the offfrequency masker case. Given the small number of data
points, neither of the correlations was significantly different
from zero. The negative correlation in the on-frequency case
was expected, but the correlation was expected to be near 0
for the off-frequency case.
PF fits to the individual data for select conditions are
shown in Fig. 10. As noted in experiments 1 and 2, there is
variability in PF slopes among the subjects. For example,
PFs for subjects S4 and S6 reflect the expected trends, that is,
PFs are steeper in the off-frequency masker conditions across
masker level in comparison to the shallower PFs in the onfrequency masker conditions. However, subject S8’s PFs are
of similar slope across all conditions. Subject S10’s PFs are
shallow across conditions, which is similar to experiment 2
in which the PFs for the same subject were shallow regardless of masker frequency, masker level, or signal delay !see
Fig. 8". There are also concerns regarding individual fits.
Although the PF fit for the off-frequency masker, 50-dB
fixed-signal level condition for S10 met the criteria for inclusion in data analysis, for example, the line appears to
approximate only one or two data points. Alternative fits
with higher values of r 2 and better visual properties are
available, but the values of % 2 associated with those fits are
much greater.
Schairer et al.: Peripheral nonlinearity and psychometric functions
TABLE III. PF thresholds and slopes, and r 2 values for PF fits for each
subject in each condition in experiment 3.
Subject
Masker
frequency
Signal
level
Threshold
Slope
r2
S4
On
30
50
57
63
25
30
40
50
30
50
57
63
25
30
40
50
30
50
57
63
25
30
40
50
30
50
57
63
25
30
40
50
27.32
54.53
75.77
79.33
61.71
65.93
72.79
77.78
26.42
51.42
60.96
77.33
32.88
64.39
72.70
80.41
26.89
50.97
58.44
67.86
53.33
59.60
65.84
71.88
21.69
43.65
59.20
68.83
58.45
67.40
73.97
77.93
1.00
0.55
0.32
0.51
1.75
1.15
1.51
0.99
0.62
0.33
0.26
0.26
0.11
1.38
0.51
1.13
0.74
1.18
0.46
0.45
0.55
0.64
1.09
0.92
0.52
0.40
0.20
0.41
0.08
0.33
0.19
0.27
1.00
0.99
0.94
0.89
1.00
0.99
1.00
0.96
0.97
0.85
0.88
0.82
0.28
0.98
0.71
0.99
0.90
0.90
0.91
0.81
0.82
0.89
0.97
0.99
0.90
0.84
0.90
0.99
0.27
0.86
0.77
0.80
Off
S6
On
Off
S8
On
Off
S10
On
Off
V. DISCUSSION
A. Methodological issues
Our confidence in the interpretation of the current PF
data is limited by several issues that are ongoing topics of
debate in the literature. The first is the use of adaptiveprocedure data to estimate PF parameters. Simulation studies
suggest that procedures with more than two intervals, combined with a decision rule that estimates a higher point on the
PF, provide more efficient and more accurate estimates of the
PF slope !e.g., Leek et al., 1992; Leek, 2001". Dai !1995",
using a similar approach based on extensive simulations,
found that estimates of PF parameters were unbiased when
obtained using the data collection and analysis procedures
we have used, although he did not explore the alternative
decision rules tested by Leek et al. !1992". While Dai !1995"
evaluated the accuracy of both the PF slope and the PF
threshold estimate, only the accuracy of the slope estimate is
of any concern here. The PF threshold estimates are highly
correlated with the estimates obtained by taking the mean of
the reversal points in the adaptive procedure (r"0.99 for the
231 thresholds in experiment 2", a widely used threshold
measure.
The second issue is the choice of equation used to fit the
PFs and to estimate the slope parameter. The estimates of PF
slope reflect the representation of the stimuli used in fitting
the PFs, as well as the assumed form of the PF. Any of the
approaches used in the detection literature would show shallower slopes at higher signal levels, but the outcome of a
statistical analysis could be somewhat dependent on the measure of slope that was used. We converted signal levels to
power and fitted the data using the same equation used to fit
PFs for absolute thresholds !e.g., Watson et al., 1972". This
approach seemed most consistent with the literature, given
that the data were obtained in a masking procedure where we
had no valid measure of the effective level of the masker at
the time the signal was presented. The Plack and Oxenham
!1998" model assumes, however, that the task is equivalent to
intensity discrimination where the decision variable is an internal representation of &I/I or (I&&I)/I. Those two quantities produce identical threshold predictions when using the
model, but they do not result in identical PFs. Use of &I/I,
FIG. 10. Example PF fits to the individual data for variable-masker conditions in experiment 3. Data for the onfrequency masker, 57- and 63-dB
fixed-signal-level conditions are represented by open triangles and squares,
respectively. Data for the offfrequency masker, 30- and 50-dB
fixed-signal-level conditions are represented by filled triangles and squares,
respectively. PF fits to the data are represented by solid lines. PFs for subjects S4 and S6 demonstrate steeper
PFs in the off-frequency masker conditions across masker level, and shallower PFs in the on-frequency masker
conditions. However, PF slopes are
similar across conditions for subjects
S8 and S10s. For S10, PFs are shallow
across conditions for variable-masker
conditions !this figure" and variablesignal conditions !see Fig. 8".
J. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
Schairer et al.: Peripheral nonlinearity and psychometric functions
1569
where I was a constant dependent on masker level, would
rescale all points on the function leaving the slope unchanged and thus would produce slope estimates identical to
those we have obtained. Conversion to (I&&I)/I would be a
nonlinear transform, with unclear results, but it seems unlikely that the differences in the fits would be large. We
assumed a form of the PF !e.g., Egan et al., 1969" that has
been widely used in the detection literature, but many alternatives have been proposed. There are only two other studies
in the literature that we are aware of that have reported PFs
for forward masking conditions. Werner !1999" plotted PFs
in percent yes votes versus signal level in dB, but did not fit
an equation to the data. Nizami and Schneider !1999" used
probit analysis to fit PFs as normal ogives in percent correct
versus signal level in dB. Their PFs were similar in form to
those generated with the Egan et al. equation.
The quality of some of the individual PF fits also is an
issue. In order to avoid ambiguity in the analyses and the
potential for selective data trimming when fitting a large
number of PFs, we adopted strict rules for removing data
points before fits were done and for excluding unreliable fits
when computing mean slope values. The criterion of r 2
#0.50 used for inclusion of the PF slope in summary data
was somewhat arbitrary, but use of other criterion points did
not have a strong effect on the correlation of slope with
signal level or masker level. Some fits appeared to be less
than optimal when the PFs were plotted with the actual data
points !e.g., Fig. 10, S10, off-frequency masker, 50-dB signal
level data" because the maximum likelihood procedure !Dai,
1995" assigned more weight to points higher on the PF.
Finally, it is a concern that the individual data were
highly variable across subjects and conditions. For example,
subject S10 demonstrated shallow PFs as a function of signal
level, regardless of masker frequency, masker level, signal
delay, or varied parameter !see Fig. 8 and Fig. 10". We do not
have an explanation for this trend. One possibility is that this
subject has greater internal noise in comparison to the other
subjects. If this is the case, the difference in internal noise
did not extend to detection of the signal in quiet, where results are in line with the remaining subjects !see the data
points represented by stars in Fig. 8". Another example of
individual variability is that a few of the subjects demonstrated higher PF slopes than would have been expected,
given the associated signal levels, in the on-frequency
masker, 0-ms signal delay, variable-signal condition !see S4,
S5, and S8 in Fig. 8". There is not sufficient information to
determine whether the slopes or the thresholds were unusually high. It also cannot be determined at this point whether
this threshold/slope effect is one that should be taken into
account in a model of forward masking, or could be attributed to a phenomenon such as confusion !Neff, 1985, 1986"
that is outside the framework of the model. A confusion effect seems unlikely given that such effects are not generally
found for sinusoidal maskers and signals, and tend to be
associated with nonmonotonic PFs, with sharp decreases in
percent correct in regions where the signal and masker levels
are nearly equal. Of the individual PFs that met the criteria
for inclusion in the mean, only one function !subject S4,
90-dB masker condition, see Fig. 7" displayed this trend.
1570
J. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
B. Effects of the peripheral nonlinearity
Despite the methodological considerations discussed in
the previous section, the peripheral nonlinearity is clearly
reflected in both the GOM and PF data. In general, GOM
functions for on- and off-frequency maskers and with various
signal delays were consistent with previous reports in the
literature for variable-signal conditions !Fastl, 1979; Jesteadt
et al., 1982; Kidd and Feth, 1981; Luscher and Zwislocki,
1947; Munson and Gardner, 1950; Oxenham and Plack,
2000; Plack and Oxenham, 1998; Widin and Viemeister,
1979; Zwislocki et al., 1959" and variable-masker conditions
!Nelson et al., 2001, 1990; Nelson and Freyman, 1987, 1984;
Oxenham and Plack, 1997". The Plack and Oxenham !1998"
model, which includes peripheral nonlinearity as a first stage,
provides an excellent account of the present GOM data for
on-frequency maskers !see the left panel of Fig. 6". The current data for off-frequency maskers show a larger effect of
signal delay on GOM functions than observed by Oxenham
and Plack !2000", but the effect was still smaller than that
observed with on-frequency maskers, in agreement with their
model.
Plack and Oxenham !1998" make no specific predictions
concerning PFs, but the assumptions of the model suggest
that PF slopes should reflect the nonlinearity, and the model
and supporting data specify the degree of nonlinearity expected for the on- and off-frequency maskers used here. The
model assumes that internal noise is Gaussian in power with
a standard deviation that is proportional to the mean !i.e.,
constant in dB", except in conditions where the masked
threshold is near quiet threshold. Given that internal noise is
constant, the slope or exponent of the PF should be proportional to the degree of nonlinearity. To a first approximation,
the results of the present experiments were consistent with
this prediction. In the variable-signal conditions of experiments 1 and 2, PF slopes decreased as a function of signal
level for on- and off-frequency maskers. This was expected
because in variable-signal conditions, the signal passes
through a range of input levels to the nonlinearity at the
place of the signal during threshold estimation. Thus, the
associated PF slopes should be influenced by the nonlinearity
at the place of the signal, regardless of masker frequency.
Given this trend, it also was expected in experiment 2 that
PF slope would be more highly correlated with signal level
at threshold than with the other variables !i.e., masker level,
signal delay, and masker frequency". However, the correlation of PF slope with masker level was greater than the correlation with signal level. This relationship is explored further in the following section.
In the variable-masker conditions of experiment 3, it
was expected that PF slope would be shallower at higher
masker levels for the on-frequency masker case, and that PF
slope would remain steep across all masker levels for the
off-frequency masker case. This is because in the variablemasker condition, masker level passes through a range of
input levels to the nonlinearity at the place of the signal as
the PF is obtained for any given condition. Thus, the onfrequency masker will be influenced by the nonlinearity at
the place of the signal, and the associated PF slopes will
reflect the amount of compression, whereas the off-frequency
Schairer et al.: Peripheral nonlinearity and psychometric functions
masker will grow linearly at the place of the signal, and the
associated PF slopes should be similar over the entire range
of masker levels. Only two of the four subjects demonstrated
this trend. Because on- and off-frequency masker cases were
examined separately, there were fewer data points available
for the correlation in experiment 3 in comparison to experiment 2 in which on- and off-frequency data were combined.
Given the small number of data points, we cannot draw any
meaningful conclusions from the correlation coefficients for
experiment 3.
Both of the previous studies that describe PFs for
forward-masked conditions used variable-signal procedures.
Werner !1999" used a 65-dB SPL broadband noise to mask a
1-kHz tone with signal delays varying from 5 to 200 ms in
normal-hearing adults and infants 3 to 6 months of age. The
PFs for the adult group were steep and parallel across signal
delay, and all PFs were restricted to signal levels below approximately 40 dB SPL. Thus, the PFs for the adult group in
the Werner study were not in a signal-level range that would
have reflected the effects of the nonlinearity as in the current
study. Even if the PFs had extended into a higher signal
range, it is not clear that the effects of the nonlinearity should
be evident in conditions using a broadband masker rather
than a pure tone masker. Nizami and Schneider !1999" obtained their data using a 97-dB, on-frequency, 2-kHz tone to
mask a Gaussian-shaped signal, and manipulating signal delay in very small steps. While we found a large effect of
masker level on PF slope in experiment 2, in the Nizami and
Schneider study, masker level was constant. Their data point
to a strong effect of signal level. They also observed fine
structure and nonmonotonicity in their PF parameters, characterized by sharp increases in masked threshold and slope
with successive increases in signal delay. This fine structure
was not observable in the current data because of the large
differences in signal delays !2, 20, and 40 ms in experiment
1, and 0, 10, and 30 ms in experiment 2 in the current study,
as opposed to 0.5-ms steps in some cases in the Nizami and
Schneider study". The general trend in their data was for
masked threshold to decrease and PF slope to increase as
signal delay increased. This is consistent with the current
data !experiments 1 and 2, on-frequency masker" and with
the interpretation that signal level at threshold determines the
slope of the PF.
The Plack and Oxenham !1998" model can be extended
to make specific quantitative predictions about PF slope ratios, in addition to the qualitative predictions reviewed
above. If the slope or exponent of the PF is truly proportional
to the degree of nonlinearity, then the compression ratio assumed in the model should be reflected in the slopes estimated from the data. Plack and Oxenham !1998" assumed a
compressive nonlinearity with a lower slope of 0.78 dB/dB
and an upper slope of 0.16 dB/dB, a ratio of 4.9 to 1 in the
model fits to their on-frequency masker data. This is similar
to the ratio of 5.2 to 1 that was obtained with the parameters
estimated from the current GOM data !see Fig. 6, lower
slope of 0.62 and upper slope of 0.12" and predicts a ratio of
5 to 1 in the slopes of the PFs. The observed ratio of the
mean PF slope at 25.8 dB (slope"0.99), the lowest signal
level in the mean data panel of Fig. 7, to the mean slope at
J. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
85.1 dB (slope"0.31), the highest signal level, was 3.2 to 1.
The smaller ratio may reflect errors in fitting the PFs.
Finally, regardless of the strength of the relationship between signal level and PF slope, the pattern of PF slopes
across conditions should predict the general form of the nonlinearity. The assumption in the Plack and Oxenham !1998"
model that the nonlinearity can be approximated by a twopart function defined by two straight lines predicts two
groups or clusters of data points in Figs. 4 and 8, one at high
PF slope values !i.e., to reflect the linear portion of the function" and the other at low slope values !to reflect the compressive portion of the function". The PF slope data suggest a
nonlinearity with a continuously decreasing slope or a very
gradual transition, consistent with more recent models that
include a peripheral nonlinearity !e.g., Glasberg and Moore,
2000".
The degree of peripheral nonlinearity has been estimated
from masking data using the comparison of slopes of GOM
for on- and off-frequency maskers in variable-masker conditions !Oxenham and Plack, 1997", from the ratio of
temporal-masking-curve slopes for on- and off-frequency
maskers !Nelson et al., 2001", and from behavioral thresholds in on-frequency masking conditions with a range of
signal delays and masker levels, such as those shown in Fig.
6. The current data suggest that measures based on slopes of
PFs might offer another alternative that, with use of optimum
procedures for measurement of PFs, could be both efficient
and broadly applicable. The ratio of slopes of GOM functions for on- and off-frequency maskers only provides an
estimate of the compressive portion of the nonlinearity, not
of the transition or the lower, steeper portion of the function.
In contrast, the ratio of fixed-probe temporal-masking-curve
slopes for on- and off-frequency maskers provides an estimate of the transfer function for a larger range of input levels, including the lower levels !Nelson et al., 2001". However, collection of a full set of data for a range of masker
levels and signal delays is time consuming and recovery of
nonlinearity parameters for either method requires a specific
model with numerous assumptions. The range of PF slopes
from a carefully selected set of conditions could be used to
provide a more complete estimate of the nonlinearity, with
fewer a priori assumptions concerning its form.
C. Exploring the masker level effect on PF slope
By far the biggest issue in the interpretation of the current data is the strong effect of masker level on the slope of
the PF. The original hypothesis was that PF slope would have
a higher correlation with signal level than with the other
variables in experiment 2 because the signal level was varied
in the adaptive procedure to reconstruct the individual PFs
within conditions. However, it was found that PF slope was
more highly correlated with masker level than with signal
level. Interpreting the relationship between PF slope and signal level, and between PF slope and masker level, is complicated by the fact that masker level and signal level are highly
correlated with each other. We examined these relationships
further using partial correlations to obtain a measure of the
relation between PF slope and each variable, with all other
Schairer et al.: Peripheral nonlinearity and psychometric functions
1571
variables held constant. The partial correlation of PF slope
with threshold was %0.06, with masker level was %0.71,
with signal delay was 0.05, and with masker frequency was
0.28. Thus, signal level at threshold had no almost relationship to PF slope once the remaining variables were accounted for in the comparison. Masker level had the strongest relationship with PF slope, indicating that as masker
level increased, PF slope decreased with signal level held
constant. This effect is not predicted by the model described
in Fig. 1. When the data from the three subjects !S4, S5, and
S8" in the 0-ms signal delay conditions were excluded from
the analysis, the overall correlation of PF slope decreased
with masker level !from %0.89 to %0.78" and increased with
signal level !from %0.69 to %0.77". A similar trend was
noted in the partial correlations of PF slope with masker
level !a decrease from %0.71 to %0.61" and with signal level
!an increase from %0.06 to %0.10". Partial correlations of PF
slope increased with signal delay to 0.29 and with masker
frequency to 0.34. This shift in the zeroth order and partial
correlations cannot account for the masker level effect, and it
does not change the conclusion that masker level has a
strong, independent contribution to PF slope.
Three factors might contribute to the strong effect of
masker level in experiment 2. First, the relationship between
signal level at threshold and PF slope is not linear, and therefore would be poorly represented by a correlation that assumes a linear relationship between variables. The mean data
panel in Fig. 7 could be viewed as nonlinear, particularly if
the 0-ms signal delay conditions are disregarded, and the
model described in Fig. 1 predicts a nonlinear effect. We do
not have enough data points to support a nonlinear regression
analysis. Second, the correlation of PF slope with signal
level at threshold is attenuated by the measurement error
associated with the latter variable. Masker level, masker frequency, and signal delay are all independent variables whose
values are not subject to measurement error. We do not have
an appropriate measure of the reliability of the estimates of
signal level at threshold that could be used to correct for
attenuation !e.g., Guilford, 1965, p. 487". Data from experiment 3, where masker level was the dependent variable, are
too limited to provide a valid comparison.
Third, the strong effect of masker level might reflect an
increase in internal noise with an increase in overall level. In
the example provided in Fig. 1, the internal noise is assumed
to be the same across conditions as represented by lower and
upper PDFs with similar standard deviations on the Y axis.
An increase in internal noise with increases in level, corresponding to the more compressive region of the function,
would be represented by an upper PDF with a larger standard
deviation. This would result in a stronger effect of masker
level than of signal level because masker levels vary over a
wider range than signal levels. We would also predict an
effect of masker level with signal level held constant. Of the
factors considered here, the assumption that internal noise is
greater at higher levels seems best able to account for the
observed pattern of results. This assumption would have a
significant impact on procedures for fitting threshold data. It
may be possible to test the assumption by measuring the
amount of internal noise !e.g., Spiegel and Green, 1981".
1572
J. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
VI. CONCLUSIONS
Growth of masking results for on- and off-frequency
forward-masker cases, in variable-masker and variable-signal
conditions, were consistent with the literature. The mean onfrequency masker data from experiment 2 were well represented by a model described by Plack and Oxenham !1998"
that incorporates a peripheral nonlinearity as its first stage.
The model assumptions suggest that this nonlinearity also
should be reflected in the PF slopes. Mean PF slopes decreased with increases in mean signal level at threshold for
on- and off-frequency masker cases in the variable-signal
conditions of experiments 1 and 2. PF slopes were shallow
for on-frequency maskers and steep for off-frequency
maskers at high thresholds for two of the four subjects in the
variable-masker conditions of experiment 3. These results
are consistent with the hypothesis that the peripheral nonlinearity is reflected in PF slopes, particularly for experiments 1
and 2. Further data collection would be necessary to make
stronger statements about the influence of the peripheral nonlinearity on PF slopes for variable-masker conditions. PF
slopes from select conditions could be used to produce more
direct estimates of the specific form of the nonlinearity than
can be obtained by fitting current models to threshold data.
An unexpected result was the strong relationship of masker
level with PF slope, independent of signal level. This finding
might be explained by assuming that internal noise increases
with increasing stimulus level. Greater internal noise would
result in shallower PFs. Given that masker levels are higher
than the corresponding signal levels at threshold in forward
masking and that the range of masker levels is greater than
the range of signal levels, an internal noise that increased
with stimulus level would result in a correlation between PF
slope and masker level that was higher than the correlation
between PF slope and signal level. Further work will be required to determine the cause of the masker level effect.
ACKNOWLEDGMENTS
This work was supported by R01 DC00136 and by T32
DC00013. We thank Andrew Oxenham and Christopher
Plack for extensive input regarding the model and Christopher Plack for providing a copy of the MatLab code used to
fit the data shown in Fig. 6. We also thank Robert Lutfi for
his instructive comments regarding the role of internal noise
in forward masking and Huanping Dai for advice on obtaining and fitting psychometric-function data.
1
The number of signal levels excluded during the PF fitting procedure varied
among subjects, and depended, for example, upon the starting level of the
signal at the beginning of each adaptive track. Because we collected data
from three to four subjects simultaneously in experiments 2 and 3, the
starting level was the same for all subjects. A greater number of signal
levels were trimmed due to rare occurrence for subjects who had lower
thresholds because they produced correct responses at a larger number of
the initial signal levels in comparison to subjects with higher thresholds. As
threshold was approached, many of these initial signal levels were not
presented again for those with lower thresholds, but were presented repeatedly for those with higher thresholds. We counted the number of signal
levels that remained after trimming the levels that were associated with low
occurrence or less than 50% correct, and averaged these value across conditions and subjects for experiments 1 and 2. The mean number of signal
Schairer et al.: Peripheral nonlinearity and psychometric functions
levels left after the trim was 5.1 (sd"1.3) for experiment 1, and 4.5 (sd
"1.5) for experiment 2.
2
The temporal window parameters used in fitting the current data were fixed
at Tb1"4.00, Tb2"29.00, w"0.16, and Ta"3.50. Tb1, Tb2, and Ta were
the same values Plack and Oxenham !1998" used to fit their data. Although
Plack and Oxenham !1998" used a different value for w in the 1998 study,
and Oxenham !2001" found better fits after allowing Tb1, Tb2, and w to
vary, we felt the fit to our data was adequate using the original temporal
window parameters listed above.
Buus, S., and Florentine, M. !1991". ‘‘Psychometric functions for level discrimination,’’ J. Acoust. Soc. Am. 90, 1371–1380.
Dai, H. !1995". ‘‘On measuring psychometric functions: a comparison of the
constant-stimulus and adaptive up-down methods,’’ J. Acoust. Soc. Am.
98, 3135–3139.
Egan, J. P., Lindner, W. A., and McFadden, D. !1969". ‘‘Masking-level differences and the form of the psychometric function,’’ Percept. Psychophys. 6, 209–215.
Fastl, H. !1979". ‘‘Temporal masking effects: III. Pure tone masker,’’ Acustica 43, 282–294.
Glasberg, B. R., and Moore, B. C. J. !2000". ‘‘Frequency selectivity as a
function of level and frequency measured with uniformly exciting notched
noise,’’ J. Acoust. Soc. Am. 108, 2318 –2328.
Guilford, J. P. !1965". Fundamental Statistics in Psychology and Education
!McGraw-Hill, New York".
Jesteadt, W., Bacon, S. P., and Lehman, J. R. !1982". ‘‘Forward masking as
a function of frequency, masker level, and signal delay,’’ J. Acoust. Soc.
Am. 71, 950–962.
Kidd, Jr., G., and Feth, L. L. !1981". ‘‘Patterns of residual masking,’’ Hear.
Res. 5, 49– 67.
Leek, M. R. !2001". ‘‘Adaptive procedures in psychophysical research,’’
Percept. Psychophys. 63, 1279–1292.
Leek, M. R., Hanna, T. E., and Marshall, L. !1992". ‘‘Estimation of psychometric functions from adaptive tracking procedures,’’ Percept. Psychophys. 51, 247–256.
Levitt, H. !1971". ‘‘Transformed up-down methods in psychoacoustics,’’ J.
Acoust. Soc. Am. 49„2, Part 2…, 467– 477.
Luscher, E., and Zwislocki, J. !1947". ‘‘The decay of sensation and the
remainder of adaptation after short pure-tone impulses on the ear,’’ Acta
Oto-Laryngol. 35, 428 – 455.
Moore, B. C., Peters, R. W., and Glasberg, B. R. !1999". ‘‘Effects of frequency and duration on psychometric functions for detection of increments and decrements in sinusoids in noise,’’ J. Acoust. Soc. Am. 106,
3539–3552.
Munson, W. A., and Gardner, M. B. !1950". ‘‘Loudness patterns—a new
approach,’’ J. Acoust. Soc. Am. 22, 177–190.
Neff, D. L. !1985". ‘‘Stimulus parameters governing confusion effects in
forward masking,’’ J. Acoust. Soc. Am. 78!6", 1966 –1976.
Neff, D. L. !1986". ‘‘Confusion effects with sinusoidal and narrow-band
noise forward maskers,’’ J. Acoust. Soc. Am. 79!5", 1519–1529.
J. Acoust. Soc. Am., Vol. 113, No. 3, March 2003
Nelson, D. A., Chargo, S. J., Kopun, J. G., and Freyman, R. L. !1990".
‘‘Effects of stimulus level on forward-masked psychophysical tuning
curves in quiet and in noise,’’ J. Acoust. Soc. Am. 88, 2143–2151.
Nelson, D. A., and Freyman, R. L. !1987". ‘‘Temporal resolution in sensorineural hearing-impaired listeners,’’ J. Acoust. Soc. Am. 81, 709–720.
Nelson, D. A., and Freyman, R. L. !1984". ‘‘Broadened forward-masked
tuning curves from intense masking tones: delay-time and probe-level manipulations,’’ J. Acoust. Soc. Am. 75, 1570–1577.
Nelson, D. A., Schroder, A. C., and Wojtczak, M. !2001". ‘‘A new procedure
for measuring peripheral compression in normal-hearing and hearingimpaired listeners,’’ J. Acoust. Soc. Am. 110, 2045–2064.
Nizami, L., and Schneider, B. A. !1999". ‘‘The fine structure of the recovering auditory detection threshold,’’ J. Acoust. Soc. Am. 106, 1187–1190.
Oxenham, A. J. !2001". ‘‘Forward masking: adaptation or integration?’’ J.
Acoust. Soc. Am. 109, 732–741.
Oxenham, A. J., and Moore, B. C. J. !1994". ‘‘Modeling the additivity of
nonsimultaneous masking,’’ Hear. Res. 80, 105–118.
Oxenham, A. J., and Plack, C. J. !1997". ‘‘A behavioral measure of basilarmembrane nonlinearity in listeners with normal and impaired hearing,’’ J.
Acoust. Soc. Am. 101, 3666 –3675.
Oxenham, A. J., and Plack, C. J. !2000". ‘‘Effects of masker frequency and
duration in forward masking: further evidence for the influence of peripheral nonlinearity,’’ Hear. Res. 150, 258 –266.
Plack, C. J., and Oxenham, A. J. !1998". ‘‘Basilar-membrane nonlinearity
and the growth of forward masking,’’ J. Acoust. Soc. Am. 103, 1598 –
1608.
Rhode, W. S. !1971". ‘‘Observations of the vibration of the basilar membrane in squirrel monkeys using the Mossbauer technique,’’ J. Acoust.
Soc. Am. 49, 1218 –1231.
Robles, L., Ruggero, M. A., and Rich, N. C. !1986". ‘‘Basilar membrane
mechanics at the base of the chinchilla cochlea. I. Input-output functions,
tuning curves, and response phases,’’ J. Acoust. Soc. Am. 80, 1364 –1374.
Ruggero, M. A., Rich, N. C., Recio, A., Narayan, S. S., and Robles, L.
!1997". ‘‘Basilar-membrane responses to tones at the base of the chinchilla
cochlea,’’ J. Acoust. Soc. Am. 101, 2151–2163.
Spiegel, M. F., and Green, D. M. !1981". ‘‘Two procedures for estimating
internal noise,’’ J. Acoust. Soc. Am. 70, 69–73.
Watson, C. S., Franks, J. R., and Hood, D. C. !1972". ‘‘Detection of tones in
the absence of external masking noise. I. Effects of signal intensity and
signal frequency,’’ J. Acoust. Soc. Am. 52, 633– 643.
Werner, L. A. !1999". ‘‘Forward masking among infant and adult listeners,’’
J. Acoust. Soc. Am. 105!4", 2445–2453.
Widin, G. P., and Viemeister, N. F. !1979". ‘‘Intensive and temporal effects
in pure-tone forward masking,’’ J. Acoust. Soc. Am. 66, 388 –395.
Yates, G. K., Winter, I. M., and Robertson, D. !1990". ‘‘Basilar membrane
nonlinearity determines auditory nerve rate-intensity functions and cochlear dynamic range,’’ Hear. Res. 45, 203–219.
Zwislocki, J., Pirodda, E., and Rubin, H. !1959". ‘‘On some poststimulatory
effects at the threshold of audibility,’’ J. Acoust. Soc. Am. 31, 9–14.
Schairer et al.: Peripheral nonlinearity and psychometric functions
1573
Download